fix(crewai): make traceai_crewai instrumentor compatible with crewai 1.x

[SuhaniNagpal7]

Summary

traceai_crewai was written against crewai 0.4x. With the instrumentor installed on a host application that has upgraded to crewai 1.x, Crew.kickoff() raises before reaching the wrapped call — so the entire crew workflow fails the moment tracing is enabled. This PR fixes the two hard 1.x crashes plus two latent issues that surface on streaming kickoffs (crew.stream=True), with the minimum diff necessary. All four changes preserve behavior on the older crewai versions the package already supported.

Verified end to end against crewai 1.14.4: instrumented Crew.kickoff ran cleanly, spans shipped to Future AGI Observe, agent run + task execute spans rendered with usage tokens (prompt 99 / completion 20 / total 119).

What was breaking on crewai 1.x

1. agent.i18n.prompt_file no longer exists — crashes every Crew.kickoff()

In crewai 1.x, i18n was demoted from an Agent field to a module-level singleton I18N_DEFAULT exposed by crewai.utilities.i18n. The wrapper's crew_agents dict-comp reads agent.i18n.prompt_file for every agent in crew.agents, which raises AttributeError. Because the surrounding span block sets record_exception=False, set_status_on_exception=False, the exception propagates out before wrapped() ever runs. Effect: every Crew.kickoff() call crashes the moment CrewAIInstrumentor().instrument() is called.

2. Task.context defaults to _NotSpecified, not None — crashes every Crew.kickoff()

In crewai 1.x Task.context is typed list[Task] | None | _NotSpecified and defaults to the sentinel. The wrapper's if task.context else None truthiness check passes for the truthy sentinel, then the list-comp [task.description for task in task.context] raises TypeError: '_NotSpecified' object is not iterable. Same crash surface as (1) — happens before wrapped() runs.

3. crew.usage_metrics can be None on streaming kickoffs

With crew.stream=True, crewai 1.x returns a CrewStreamingOutput immediately and never reaches self.usage_metrics = self.calculate_usage_metrics() inside Crew.kickoff before returning. The wrapper's existing else branch then reads usage_metrics.prompt_tokens on None and raises AttributeError.

4. CrewStreamingOutput has no to_dict()

The wrapper's if crew_output_dict := crew_output.to_dict(): raises AttributeError when streaming returns the streaming-output wrapper instead of a CrewOutput.

What this patch changes

All scoped to python/frameworks/crewai/traceai_crewai/_wrappers.py:

• agent.i18n.prompt_file → safe helper. Added a small _agent_i18n_prompt_file(agent) that prefers agent.i18n (older crewai) and falls back to the crewai.utilities.i18n.I18N_DEFAULT singleton (1.x). The I18N_DEFAULT import is wrapped in try/except ImportError so the package still loads on very old crewai where the singleton path doesn't exist.

• Task.context truthiness → isinstance(list) guard. Now [t.description for t in task.context] if isinstance(task.context, list) else None. Also renames the inner loop variable from task to t so it stops shadowing the outer task in the surrounding comprehension — a latent bug that was just masked by the earlier crash.

• else: → elif usage_metrics is not None:. Streaming runs skip the object branch instead of crashing on None.prompt_tokens. The earlier isinstance(usage_metrics, dict) branch is unchanged.

• Walrus on crew_output.to_dict() → getattr + callable guard. to_dict = getattr(crew_output, 'to_dict', None); crew_output_dict = to_dict() if callable(to_dict) else None — falls through to str(crew_output) when absent.

What's NOT changed

I deliberately kept the diff to the four 1.x regressions. The following were considered and dropped as out-of-scope:

• isinstance(args[0], Agent) in _get_input_value — would degrade for non-Agent BaseAgent subclasses (LiteAgent, agent adapters), but isn't a crash on a normal CrewAI Agent. Future-proofing, not a 1.x regression.
• span.set_attribute("function_calling_llm", instance.function_calling_llm) — function_calling_llm is now an LLM object in 1.x, but OTel silently drops non-primitive attribute values. No crash; same behavior as 0.x where it was a LangChain ChatModel.
• A duplicated set_attribute(OUTPUT_VALUE, ...) line in _ToolUseWrapper. Pre-existing dead code, not in scope.
• LLM model / cost capture — by design the CrewAI instrumentor only emits framework spans (Crew.kickoff, Task._execute_core, ToolUsage._use). Model name and cost come from running OpenAIInstrumentor (or equivalent provider instrumentor) alongside this one. Same pattern as the other framework instrumentors in this repo.

Test plan

• python -m py_compile python/frameworks/crewai/traceai_crewai/_wrappers.py — syntax clean.
• Fresh venv (Python 3.12) with crewai==1.14.4, fi-instrumentation-otel==1.0.0, and traceai-crewai installed editable from this branch.
• Smoke test: single-agent + single-task Crew, crew.kickoff() against OpenAI gpt-4o-mini. Before fix: crashes with _NotSpecified TypeError. After fix: runs to completion, returns model output, no exceptions.
• Trace lands in Future AGI Observe under the configured project:
  • Crew.kickoff root span: status OK, duration 3.4s, tokens prompt=99 / completion=20 / total=119, gen_ai.span.kind=CHAIN, crew_agents and crew_tasks blobs populated, crew_tasks[].context = null (correctly handled _NotSpecified sentinel), crew_agents[].i18n = null (correctly handled missing field), output.value = the agent's generated text.
  • Task._execute_core child span present and parented correctly.

Files changed

• python/frameworks/crewai/traceai_crewai/_wrappers.py — 28 insertions, 6 deletions.